2

In the very same spirit as my previous question I am trying to find an explicit representation of the convex conjugate $f^*$ of the function $$ f \colon \mathbb R \to [0, \infty], \qquad x \mapsto | 1 - x |^{\frac{1}{a}} (1 + x)^{\frac{\alpha - 1}{\alpha}} + \iota_{[0, \infty)}(x), $$ where for any set $A$, the convex indicator function $\iota_{A}(x)$ is equal to $0$ for $x \in A$ and equal to $\infty$ else and $\alpha \in (0, 1)$.

Up to edge cases, this first boils down to finding the inverse of $$ f'(x) = \frac{1}{\alpha} (x+1)^{\frac{\alpha - 1}{\alpha}} | x - 1 |^{\frac{1}{\alpha}} \frac{\alpha(x - 1) + 2}{(x- 1)(x + 1)} = \frac{1}{\alpha} \left| \frac{x-1}{x+1}\right|^{\frac{1}{\alpha}} \left(\alpha + \frac{2}{(x- 1)}\right) $$ (to obtain $f^*(y) = y (f')^{-1}(y) - f((f')^{-1}(y))$.)

Due to the last term in parenthesis, I do not know how to invert $f'$, like it was done in this answer. We have that $f'$ is strictly monotonically increasing and continuous, so that it is bijective onto its range and hence invertible.

For $\alpha = \frac{2}{3}$, WolframAlpha gives an explicit inverse, which as a quite complicated expression, whereas for $\alpha = \frac{1}{\sqrt{2}}$, it can't find an answer in terms of standard mathematical functions.

ViktorStein
  • 4,838

1 Answers1

1

$\def\a{\alpha}\DeclareMathOperator\I{I}$

Substitute $x=\frac{w+1}{w-1}$ (i.e. $w = \frac{x + 1}{x - 1}$):

$$\frac1\a\left|\frac{x+1}{x-1}\right|^\frac1\a\left(\a+\frac2{x-1}\right)=y\iff |w|^\frac1a-\frac w{\a y}-\frac{\a-1}{\a y}=0$$

For $y < 0$ there is a closed Wolfram language form via this post in section 1

$$x^r+ax+b=0\implies x=\frac{b r}{a(1-r)\text I^{-1}_\frac{b^{r-1}(r-1)}{a^r\left(\frac1r-1\right)^r}(r-1,2)}$$

where inverse beta regularized appears. Therefore:

$$\boxed{f(x)=|1-x|^\frac1\a(1+x)^\frac{\a-1}\a\implies (f’)^{-1}(x)=\frac2{\I^{-1}_{1-x}\left(2,\frac1\alpha-1\right)}-1,0\le x\le1}$$

testable here. Wolfram Alpha automatically reduces the solution to a radical form if possible.

Using $(f’)^{-1}(y)=\frac2u-1$, then $f^*(y)=2\frac{y-|1-u|^\frac1a}u-y$ using algebra and software. $1-u$ from section 7 can be used again and the absolute value removed as $u\ge0$ to get $$\boxed{f^*(y)=2\frac{y-\I^{-1}_y\left(\frac1\a-1,2\right)^\frac1\a}{\I^{-1}_{1-y}\left(2,\frac1\a-1\right)}-y,0\le y\le 1}$$

There is probably also a series solution, with closed forms for possibly the entire domain of $f(x)$ which will be added later.

Тyma Gaidash
  • 12,081
  • The result is for the branch of the inverse in all of quadrant $1$. However, the absolute value makes it hard to come up with an $I^{-1}_x(a,b)$ solution for the other branch – Тyma Gaidash Dec 24 '23 at 14:30
  • @ViktorStein There should have been a more detailed derivation, but you can use $I^{-1}x(a,b)=1-I^{-1}{1-x}(b,a)$ (section 7 of the linked post) and find $x=\frac{w+1}{w-1}=\frac{\frac1{1-I}+1}{\frac1{1-I}-1}=\frac2I-1,I=I^{-1}_z(1-\frac1\alpha,2)$ – Тyma Gaidash Dec 26 '23 at 22:14
  • Does one need $a > 0$ for the formula from section 1? – ViktorStein Feb 11 '24 at 15:36
  • @ViktorStein Yes one does. For the formula in this answer, $r-1> 0$, so $r>1$ – Тyma Gaidash Feb 11 '24 at 15:38
  • Because in section 1 it says "All three formulas account for any real $r$". – ViktorStein Feb 11 '24 at 15:44
  • @ViktorStein There are a few formulas for $-r>0,r-1>0,\frac1r-1>0$, while $r=0,1$ is trivial and likely a limit can be used for these cases. Thus, $r\in\Bbb R$ is covered. – Тyma Gaidash Feb 11 '24 at 15:51
  • Ah, I see, thank you. – ViktorStein Feb 11 '24 at 15:57